I build LLM systems and measure whether they actually work: serving stacks, performance studies, RAG and agents, evaluation pipelines, and fine-tuning experiments—including the ones that fail.

LinkedIn · Email · Israel

• Diagnosed why 4× H100 DDP was 2.4× slower than one H100: 93.5% of the representative step was gradient synchronization after NCCL fell back to sockets.
• Increased 7B serving throughput 49% with FP8 while reducing logged model weight memory 42.8% on an H100.
• Accelerated an autoregressive decode loop 5.11× and a 256-step decode 4.30× with a manually captured CUDA graph.
• Reduced a text-to-SQL agent's P95 latency from 98.5 s to 3.44 s at roughly 10 RPS without reducing the measured evaluation pass rate.
• Built reproducible RAG and coding-agent evaluations that preserve configuration, predictions, trajectories, metrics, and unsuccessful experiments.

A measured 1× vs 4× H100 study of GPT-2 Large on Kubernetes. The negative scaling result is traced to a 2.88 GiB gradient payload and NCCL socket fallback, while evaluation throughput still scales 3.84×.

PyTorch DDP · H100 · NCCL · Kubernetes · SkyPilot

Quantization from first principles plus an end-to-end BF16/FP8 vLLM benchmark. FP8 delivered 7,159 output tok/s versus 4,805 tok/s in BF16 and lowered median inter-token latency by 31.3%.

vLLM · FP8 · GuideLLM · H100 · PyTorch

Four case studies covering roofline analysis, KV-cache decode optimization, torch.compile, CUDA graphs, and speculative decoding. Includes the cases where default compilation was slower than eager execution.

CUDA · torch.compile · CUDA Graphs · KV Cache · Speculative Decoding

Qwen3-30B-A3B text-to-SQL serving on one H100 with vLLM, LangGraph, Prometheus, Grafana, and Langfuse. Iterative load testing brought the full agent path under a five-second P95 SLO.

vLLM · LangGraph · Prometheus · Grafana · Langfuse

An Airflow/MLflow pipeline for running mini-swe-agent on SWE-bench, evaluating patches, and producing self-describing run directories. Verified end to end on a small three-instance live run; not presented as a model-quality benchmark.

SWE-bench · Airflow · MLflow · Docker · MLOps

Eight controlled experiments over 42 SEC filings. Prompt policy improved answer correctness 48%, metadata filtering improved page-hit@1, and both reranking and stricter grounding produced useful negative results.

LangChain · FAISS · RAGAS · BGE · BM25 + RRF

• Glass-Box PPO — PPO/RLHF mechanics written without PPOTrainer, including GAE, clipping, KL, and a concrete reward-hacking failure.
• Customer Service Analyst Agent — typed LangGraph tools, persistent memory, recovery from bad tool arguments, and an MCP server.
• LLM-as-a-Judge Evaluation — human/judge agreement analysis showing one-directional leniency and why deterministic constraints should stay in code.
• Level-Adaptive QLoRA — an explicitly labeled negative-results study preserving the dataset, compact adapters, failure analysis, and a valid GPU rerun protocol.

LLM/ML: Python · PyTorch · Transformers · vLLM · LangGraph · LangChain · FAISS · PEFT/LoRA · MCP

Performance/MLOps: CUDA · DDP/NCCL · Docker · Kubernetes · Airflow · MLflow · Prometheus · Grafana · Langfuse

Applications/Data: FastAPI · React · TypeScript · PostgreSQL · Redis · pandas

AI / ML Engineer · LLM Engineer · ML Performance Engineer roles where measured system behavior matters as much as the demo.

Chess ♟️ · mobile gaming 📱